A drug and disease association prediction method based on hyperbolic graph feature learning network

Abstract

The application discloses to the technical field of drug and disease association prediction, specifically a drug and disease association prediction method based on a hyperbolic graph feature learning network, including the following specific steps: S1: obtaining drug-disease association data from a database, for constructing a drug similarity matrix DR and a disease similarity matrix DS, mapping the extracted features to a hyperbolic space to obtain a hyperbolic drug initial feature matrix ZR H , a hyperbolic disease initial feature matrix ZD H , for constructing a drug-disease adjacency matrix A. The application effectively extracts similar features and heterogeneous features through a hyperbolic graph feature reconstructor and a hyperbolic heterogeneous variation graph converter, improving the accuracy of the features. In addition, the hyperbolic collaborative representation learning strategy effectively fuses the similar features and the heterogeneous features, improving the comprehensiveness of the features. Furthermore, the positive-negative fusion difficult case sampling strategy synthesizes the most informative negative samples, which can better distinguish positive and negative samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug-disease association prediction, in particular to a drug-disease association prediction method based on a hyperbolic graph feature learning network. BACKGROUND

[0002] Drug repositioning (DR), also known as drug repurposing, is a promising approach in the field of pharmaceutical innovation. It discovers new clinical applications for approved drugs in a fast and cost-effective manner, thus attracting widespread attention from academia. The traditional drug discovery process usually takes 10 to 17 years and costs from 500 million to 2 billion US dollars, and it takes at least ten years to bring a drug to market, which faces many challenges. Especially in the context of the rapid development of artificial intelligence, drug repositioning, as an effective strategy to accelerate drug discovery, is gradually becoming mainstream. It discovers new indications by reusing approved drugs, which not only saves a lot of time and cost, but also reduces the risk of high failure rates. In all stages of drug discovery, the identification of new indications for existing drugs is crucial. In recent years, computational methods based on biomedical data have made significant progress in drug repositioning. These methods extract features from biochemical and medical data related to drugs and diseases, use machine learning (ML) and deep learning (DL) models to predict drug-disease associations (DDA), and thus accelerate the drug discovery process. However, many methods simply concatenate data from different modalities to integrate information, which limits their ability to capture features from a comprehensive and in-depth perspective. Therefore, innovative computational methods need to better integrate multi-modal information to improve the efficiency and accuracy of drug repositioning and further promote the progress of drug discovery.

[0003] As a deep learning (DL) method, graph neural network (GNN) is good at processing graph-structured data and can effectively capture the complex relationships and topological structures between nodes in the drug-disease network, thus having great potential in predicting the potential associations between drugs and diseases. However, existing GNN-based drug-disease association prediction models mostly learn drug and disease features in Euclidean space. For drug-drug homophily graphs and disease-disease homophily graphs, linear transformations and distance metrics in Euclidean space will oversimplify the similarity relationships between nodes, especially when the data presents a power-law distribution or significant hierarchical features. This simplification will prevent the model from accurately capturing complex association patterns between nodes, ultimately affecting the accuracy of drug-disease association prediction. While drug-disease heterophily graphs usually present a tree structure, the volume of which increases exponentially with the radius, while the volume of a sphere in Euclidean space grows polynomially, which is prone to distortion,

[0004] In addition, negative samples play a crucial role in drug and disease association prediction, but existing methods usually train by randomly sampling negative samples. This approach may overlook important negative samples or introduce irrelevant noise samples, which in turn affects model performance. Therefore, a drug and disease association prediction method based on a hyperbolic graph feature learning network is proposed. This method not only comprehensively understands the complex relationship between drug and disease nodes in the graph in hyperbolic space, but also synthesizes challenging negative samples, thereby further improving the accuracy of drug and disease potential association prediction. This method is not only innovative in theory, but also has important value in practical application.

[0005] Based on the above, the existing drug and disease association prediction model based on graph convolution network has the following problems:

[0006] 1. Existing drug and disease association prediction models usually learn similar features of drug nodes and disease nodes in homogeneous graphs in Euclidean space. However, in Euclidean space, linear transformation and distance measurement oversimplify the similarity relationship between nodes. Especially when the data presents a power-law distribution or significant hierarchical features, this simplification will lead the model to fail to accurately capture the complex association patterns between homogeneous graph nodes, thus failing to learn high-quality similar features of drug nodes and disease nodes, ultimately affecting the accuracy of drug-disease association prediction.

[0007] 2. Existing drug and disease association prediction models usually learn heterogeneous features of drug nodes and disease nodes in heterogeneous graphs in Euclidean space. However, drug and disease heterogeneous graphs usually contain hierarchical structure, and the volume of the graph increases exponentially with the radius, while the volume of the sphere in Euclidean space increases polynomially with the radius. This difference makes Euclidean space prone to distortion when modeling, i.e., the geometric properties of Euclidean space do not match the hierarchical and tree-like structure of heterogeneous graphs, leading the model to fail to accurately capture the semantic association and topological relationship between nodes, which are important components of heterogeneous features. They are crucial for understanding the complex relationship between drug nodes and disease nodes. Therefore, the limitations of Euclidean space will affect the model's ability to predict drug and disease associations.

[0008] 3. Limited feature representation capability. Although deep learning methods have improved feature extraction through powerful representation learning capabilities, many methods still fail to fully integrate the similar features of drug nodes and disease nodes and their heterogeneous characteristics, resulting in the model being unable to generate comprehensive and high-quality feature representations.

[0009] 4. The importance of negative samples in drug and disease association prediction tasks has been widely recognized, but how to select the most representative negative samples remains a challenge. Existing methods usually train by randomly extracting negative samples, but this method ignores important negative samples or introduces irrelevant noise samples, affecting the performance of the model.

[0010] Therefore, a drug and disease association prediction method based on a hyperbolic graph feature learning network is invented. SUMMARY

[0011] To solve the above technical problems, according to one aspect of the present application, the present application provides the following technical solutions:

[0012] A drug and disease association prediction method based on a hyperbolic graph feature learning network includes the following specific steps:

[0013] S1: Obtain drug-disease association data from a database for constructing a drug similarity matrix DR and a disease similarity matrix DS, and map the features extracted therefrom to a hyperbolic space to obtain a hyperbolic initial drug feature matrix ZR H , a hyperbolic initial disease feature matrix ZD H , for constructing a drug-disease adjacency matrix A;

[0014] S2: Apply a hyperbolic graph feature reconstructor to ZR H and ZD H to obtain a hyperbolic drug similarity feature matrix SR H and a hyperbolic disease similarity feature matrix SD H ;

[0015] S3: Construct a heterogeneous graph AG according to A, and train ZR H and ZD H based on AG using a GCN and a hyperbolic heterogeneous variational graph transformer to obtain a hyperbolic drug heterogeneous feature matrix HR H and a hyperbolic disease heterogeneous feature matrix HD H ;

[0016] S4: splice SR H and HR H , splice SD H and HD H , and use an MLP to obtain a hyperbolic drug feature matrix FR H and a hyperbolic disease feature matrix FD H ;

[0017] S5: Based on FR H and FD H , use a positive-negative fusion hard example sampling strategy to construct a hyperbolic hard example drug feature matrix FR' H and a hyperbolic hard example disease feature matrix FD'H ;

[0018] S6: outputting a prediction result by the MLP using FR' H and FD' H .

[0019] As a preferred scheme of the drug-disease association prediction method based on a hyperbolic graph feature learning network, the specific steps of S1 are as follows:

[0020] S11: constructing a drug similarity matrix DR and a disease similarity matrix DS, and obtaining a drug initial feature matrix ZR and a disease initial feature matrix ZD by the MLP;

[0021] S12: mapping the drug initial feature matrix ZR and the disease initial feature matrix ZD to a hyperbolic space;

[0022] S13: constructing a drug-disease adjacency matrix A

[0023] As a preferred scheme of the drug-disease association prediction method based on a hyperbolic graph feature learning network, the specific steps of S11 are as follows:

[0024] S111: obtaining drug-disease association data from a database, constructing a drug similarity matrix DR and a disease similarity matrix DS, and obtaining a drug initial feature matrix ZR and a disease initial feature matrix ZD by the MLP; wherein nr represents the number of drug nodes, and nd represents the number of disease nodes;

[0025] S112: performing feature extraction on DR and DS by the MLP to obtain a drug initial feature matrix ZR and a disease initial feature matrix ZD, and the formula is as follows:

[0026] ZR=MLP(DR)

[0027] ZD=MLP(DS)

[0028] wherein, ZR represents the drug initial feature matrix, d0 is the feature dimension thereof, ZR represents the i-th row drug initial feature vector of ZR, i.e., the drug initial feature vector of the drug node dr i . ZD represents the disease initial feature matrix; ZD represents the i-th row disease initial feature vector of ZD, i.e., the disease initial feature vector of the disease node ds i .

[0029] The specific steps of S12 are as follows:

[0030] ​S121: adopt the Lorentz model, let K=1, then is expressed as the origin, and then is expressed as

[0031] S122: for ZR i , map it to hyperbolic space, the formula is as follows:

[0032]

[0033] where, represents the hyperbolic drug initial feature vector of the drug node dr i , exp o (·) acts to map the variable to the hyperbolic space ; (0, ZR i ) satisfies represents the Minkowski inner product; ZR i represents the drug initial feature vector of the drug node dr i ; ||.||2 represents the L2 norm;

[0034] S123: by performing the above operation on the drug initial feature vector ZR i of any drug node dr i , a hyperbolic drug initial feature matrix

[0035] S124: repeat the above steps to obtain a hyperbolic disease initial feature matrix

[0036] The specific steps of the S13 are as follows:

[0037] Obtain drug-disease association data and construct an adjacency matrix A, and the specific construction method is as follows:

[0038]

[0039] As a preferred scheme of the drug-disease association prediction method based on the hyperbolic graph feature learning network, wherein the specific steps of the S2 are as follows:

[0040] S21: extract the intermediate feature vectors of the drug nodes and the disease nodes by using hyperbolic feature transformation;

[0041] S22: calculate the hyperbolic homogeneous fusion attention, and extract the hyperbolic similar features of the drug and disease nodes using the hyperbolic homogeneous fusion attention.

[0042] As a preferred scheme of the drug and disease association prediction method based on the hyperbolic graph feature learning network, in the S21, the specific steps are as follows:

[0043] S211: For the nr drug nodes in the drug-disease association data, the hyperbolic drug initial feature matrix ZR H ; represents the i-th row of the hyperbolic drug initial feature vector of ZR H ;

[0044] In the 0th layer, the hyperbolic drug similar feature matrix is as follows:

[0045] SR 0,H = ZR H

[0046] Then represents the i-th row of the hyperbolic drug similar feature vector of SR 0,H ;

[0047] S212: For any drug node dr i in the drug-disease association data, in the lth layer hyperbolic graph feature reconstructor, first, the hyperbolic feature transformation is performed to map the hyperbolic space of the last layer to the hyperbolic space of the layer, and the hyperbolic feature transformation formula of the lth layer is as follows:

[0048]

[0049] wherein, represents the intermediate feature vector of the drug node dr i in the hyperbolic space after the hyperbolic feature transformation; is the trainable parameter of the lth layer, d l is the dimension of the lth layer; is the hyperbolic drug similar feature vector of the drug node dr i in the l-1th layer; b l is the Euclidean vector of the lth layer in , which is a trainable parameter; represents the intermediate variable of the drug node dr i in the hyperbolic space of the next layer l; exp o (·) is used to map the variable to the hyperbolic space of the next layer l; log o (·) is used to map the variable to the tangent space of the origin o; is used to map the variable to the hyperbolic space with the point as the center; b l Tangent space from origin o Parallel transport to point Tangent space from

[0050] The specific steps of S22 are as follows:

[0051] S221, calculate hyperbolic homogeneous fusion attention, which is composed of node feature attention, hyperbolic distance attention and structure attention;

[0052] S2211: calculate the node feature attention of drug node dr i in the lth layer; j The specific formula is:

[0053]

[0054] Wherein, LeakReLU is an activation function; is a trainable parameter; log o The function of (·) is to map the variable to the tangent space from origin o Indicates and Perform matrix multiplication.

[0055] Then normalize to get the final node feature attention The formula is as follows:

[0056]

[0057] Wherein, Indicates the first-order neighbor set of drug node dr i , including itself; the higher the similarity of drug node dr i and drug node dr j in the node feature of the lth layer, The larger the value is;

[0058] S2212: calculate the hyperbolic distance attention of drug node dr i in the lth layer; j The specific formula is:

[0059]

[0060] Wherein, δ and η are hyperparameters, used to adjust the hyperbolic distance; Used to calculate the hyperbolic distance between and ​​​denotes the square of hyperbolic distance;

[0061] Then normalization is performed to obtain the final hyperbolic distance attention The specific formula is:

[0062]

[0063] wherein, denotes the first-order neighbor set of the drug node dr i , including itself; when the drug node dr i and the drug node dr j are closer in the l-th layer, the hyperbolic distance is smaller, and the structural attention is larger;

[0064] S2213: Calculate the structural attention of the drug node dr i to the drug node dr j , and the specific calculation process is as follows:

[0065] Given the drug similarity matrix DR, the degree of the drug node dr i is defined as:

[0066]

[0067] wherein, the formula means summing each column j of the i-th row of DR, that is, summing the similarity of the drug node dr i to each drug node dr j ;

[0068] The structural attention of the drug node dr i to the drug node dr j The calculation formula is as follows:

[0069]

[0070] wherein,

[0071] Then normalization is performed to obtain the final structural attention The formula is as follows:

[0072]

[0073] wherein, denotes the first-order neighbor set of the drug node dr i , including itself; when the drug node dr i and the drug node dr j are closer in the l-th layer, the hyperbolic distance is smaller, and the structural attention is larger;

[0074] ​S2214: In the first layer, the node feature attention is obtained by Hyperbolic distance attention And structural attention The hyperbolic homogeneous fusion attention is obtained by fusion, and the formula is as follows:

[0075]

[0076] Where w NF , w DS , and w ST are trainable parameters, all initialized to 1;

[0077] S222: Hyperbolic similar features of drug and disease nodes are extracted using hyperbolic homogeneous fusion attention;

[0078] S2221: Hyperbolic encoding is performed on to obtain the drug node dr i Hyperbolic similar features in the first layer The formula of hyperbolic encoding is as follows:

[0079]

[0080] Where, The function is to map the variable to the hyperbolic space ; σ is an activation function; N i represents the first-order neighbor set of the drug node dr i , including itself; The function is to map the variable to the tangent space of ;

[0081] S2222: The above operation is performed on any drug node dr i After h layer iterations, the hyperbolic drug similar feature matrix is finally obtained. SR h,H is represented by

[0082] S2223: The above steps are repeated to obtain the hyperbolic disease similar feature matrix

[0083] As a preferred scheme of the drug and disease association prediction method based on the hyperbolic graph feature learning network, the specific steps of S3 are as follows:

[0084] S31: Hyperbolic disease heterogeneous initial feature matrix and hyperbolic drug heterogeneous initial feature matrix are obtained using GCN;

[0085] S32: Calculate hyperbolic heterogeneous variation attention, and learn the heterogeneous characteristics of drug nodes and disease nodes using hyperbolic heterogeneous variation attention.

[0086] As a preferred scheme of the drug and disease association prediction method based on the hyperbolic graph feature learning network, wherein the specific steps of S31 are as follows:

[0087] S311: Construct a heterogeneous graph AG based on the adjacency matrix A, as follows:

[0088]

[0089] wherein,

[0090] S312: In order to embody the number of layers, let MR 0,H be represented as ZR H , and MD 0,H be represented as ZD H ; in order to use the GCN formula, splice MR 0,H and MD 0,H to obtain:

[0091]

[0092] wherein, E 0 is a feature matrix spliced by MR 0,H and MD 0,H in the 0th layer;

[0093] S313: In the lth layer, use GCN to obtain E l based on the heterogeneous graph AG, and the formula is as follows:

[0094]

[0095] AG'=(AG+I)

[0096]

[0097] wherein, is the i-th row hyperbolic heterogeneous initial feature vector of E l in the lth layer; σ is an activation function; exp o (·) is used to map variables to hyperbolic space is the first-order neighbor set of the drug node dr i or the disease node ds i , including itself; is the processed adjacency matrix, is the value in the ith row and jth column; I is the identity matrix; are trainable parameters; log o (·) acts to map the variable to the tangent space of the origin o D matrix is used to calculate

[0098] S314: After h layer iterations, we get:

[0099]

[0100] That is, from E h , we get the hyperbolic drug heterogeneous initial feature matrix and the hyperbolic disease heterogeneous initial feature matrix

[0101] The specific steps of S32 are as follows:

[0102] Given the hyperbolic drug heterogeneous initial feature matrix and the hyperbolic disease heterogeneous initial feature matrix, that is, MR h,H and MD h,H ;

[0103] In the 0th layer, the hyperbolic heterogeneous drug feature matrix and the hyperbolic disease heterogeneous feature matrix are respectively:

[0104] HR 0,H = MR h,H

[0105] HD 0,H = MD h,H

[0106] The hyperbolic heterogeneous variable graph converter has three inputs: the query matrix Q 1 and Q 2 , the key matrix K 1 and K 2 , and the value matrix V; if the hyperbolic drug heterogeneous feature is extracted, the query matrix is obtained through HR 0,H , and the key matrix and the value matrix are obtained through HD 0,H ; if the hyperbolic disease heterogeneous feature is extracted, the situation is just the opposite; in the lth layer, the outputs are HR l,H and HD l,H ;

[0107] In the lth layer, the process of obtaining HR l,H is as follows:

[0108]

[0109] wherein, HR l-1,H is the hyperbolic drug heterogeneous feature matrix of the (l-1)th layer; exp o (·) acts to map the variable to hyperbolic space log o (·) acts to map the variable to the tangent space at the origin o denotes and perform matrix multiplication, by analogy;

[0110]

[0111] where HD l-1,H is the hyperbolic disease heterogeneity feature matrix of the (l-1)-th layer; has the same meaning as above;

[0112] The update equation of the drug node dr i of the l-th layer is as follows:

[0113]

[0114] where, denotes the hyperbolic drug heterogeneity knowledge feature vector of the drug node dr i of the l-th layer; V j is the j-th row of V; denotes the Lorentz norm; is the hyperbolic drug heterogeneity feature vector of the i-th row of HR l,H ; Norm is LayerNorm or BatchNorm; exp o (·) acts to map the variable to hyperbolic space log o (·) acts to map the variable to the tangent space at the origin o

[0115] RDIF ij is defined as:

[0116]

[0117] where RDIF ij is the hyperbolic variable attention score, λ l is defined as:

[0118]

[0119] where λ l is a scalar of the l-th layer, is a learnable parameter of the l-th layer, and λ init is a hyperparameter;

[0120] After h-layer iteration, drug node dr i The output hyperbolic drug heterogeneous feature vector is Further, the hyperbolic drug heterogeneous feature matrix is obtained The HR H represents the HR h,H ;

[0121] Similarly, the hyperbolic disease heterogeneous feature matrix HD H ;

[0122] As a preferred scheme of the drug and disease association prediction method based on the hyperbolic graph feature learning network, the specific steps of S4 are as follows:

[0123] The hyperbolic drug and disease feature matrix is calculated, and the formula is as follows:

[0124] FR H = exp o (MLP(log o (SR H )||log o (HR H )))

[0125] FD H = exp o (MLP(log o (SD H )||log o (HD H )))

[0126] Wherein, exp o (·) is used to map the variable to the hyperbolic space log o (·) is used to map the variable to the tangent space of the origin o ; the role of MLP is to reduce the dimension of the hyperbolic drug and disease feature vectors from to

[0127] As a preferred scheme of the drug and disease association prediction method based on the hyperbolic graph feature learning network, the specific steps of S5 are as follows:

[0128] S51: Constructing drug negative samples and disease negative samples;

[0129] S52: Using positive-negative fusion hard example sampling strategy to generate hard example drug negative samples and hard example disease negative samples and obtain hyperbolic hard example drug feature matrix and hyperbolic hard example disease feature matrix;

[0130] The specific steps of S51 are as follows:

[0131] 511: First, map the hyperbolic drug feature matrix and the hyperbolic disease feature matrix to the drug feature matrix XR and the disease feature matrix XD as follows:

[0132] XR = log o (FR H )

[0133] XD = log o (FD H )

[0134] wherein, log o has the effect of mapping the variable to the tangent space of the origin o;

[0135] S512: Obtain drug-disease association data to generate a heterogeneous graph G, which contains drug-disease relationships, drug-drug relationships, and disease-disease relationships; find all positive sample pairs and negative sample pairs from the heterogeneous graph G. First, ensure that the number of positive sample pairs and negative sample pairs is the same. The strategy is to randomly select a part of the negative sample pairs from the heterogeneous graph G to ensure that the number is the same as the positive sample pairs, and vice versa. Finally, select drug negative samples and disease negative samples from the screened negative sample pairs; wherein the drug negative sample feature matrix is the disease negative sample feature matrix is N neg is the number of negative sample pairs, that is, the number of drug negative samples and disease negative samples;

[0136] The specific steps of S52 are as follows:

[0137] S521: Construct a difficult example candidate negative sample feature matrix;

[0138] For any node in the heterogeneous graph G, the set composed of all first-order and second-order neighbor nodes is called the neighbor set of the node;

[0139] For all drug negative samples, find all first-order and second-order neighbors, i.e., the neighbor set;

[0140] For any drug negative sample r, the operation of constructing a difficult example candidate negative sample feature matrix is as follows:

[0141] First, find all drug positive samples in the neighbor set according to the heterogeneous graph G, and the local neighbor drug positive sample feature matrix composed of the drug positive samples is: XR pl Then, fuse the feature vectors of all drug positive samples to obtain the drug fusion positive sample feature vector x​pr , where

[0142]

[0143] where is the drug positive sample feature vector of the mrth row of XR pl ; mrrepresents the number of drug positive samples within the neighbor set of drug negative sample r;

[0144] Similarly, the feature vectors of all disease positive samples within its neighbor set are fused to obtain the disease fusion positive sample feature vector x pd , where

[0145]

[0146] where is the disease positive sample feature vector of the mdth row of XD pl ; mdrepresents the number of disease positive samples within the neighbor set of drug negative sample r;

[0147] Then the disease negative samples of the second-order neighbors within the neighbor set of drug negative sample r are found, and the second-order neighbor disease negative sample feature matrix XD nl is obtained; the disease negative sample probability distribution of the disease negative samples of the second-order neighbors within the neighbor set of drug negative sample r is calculated according to x pr and x pd , where

[0148]

[0149] where, represents the feature vector of the qth row of XD nl ; μ is a hyperparameter for balancing the influence of drugs and diseases; N represents the number of disease negative samples of the second-order neighbors within the neighbor set of drug negative sample r; the meaning of this formula is that the closer the distance (x pr , x pd ), the greater the probability of being selected as a disease negative sample;

[0150] Then, according to the disease negative sample probability distribution , the disease negative samples of the second-order neighbors within the neighbor set of drug negative sample r corresponding to the M largest probability values are selected to construct a candidate disease negative sample feature matrix T of size M M is a hyperparameter;

[0151] Next, the information of the disease fusion positive sample feature vector x pd is injected into the candidate disease negative sample feature matrix T to construct a hard example candidate disease negative sample feature matrix T', where

[0152] T' i = αx pd + (1-α)T i ,α∈(0,1)

[0153] wherein T' i is the i-th row of T' of the difficult example disease negative sample feature vector; α is a hyper parameter; T i is the i-th row of T of the candidate disease negative sample feature vector;

[0154] S522: Obtain a hyperbolic difficult example drug feature matrix and a hyperbolic difficult example disease feature matrix;

[0155] For each row of the difficult example candidate disease negative sample feature matrix T', the difficult example disease negative sample feature vector T' j is obtained by performing dot product with the drug negative sample feature vector of the drug negative sample r, and the formula is as follows:

[0156]

[0157] wherein, is the difficult example disease negative sample feature vector with the largest dot product selected from T';

[0158] Since there are N neg drug negative samples, N neg difficult example disease negative sample feature vectors can be generated to form a difficult example disease negative sample feature matrix

[0159] Similarly, for the disease negative sample, N neg difficult example drug negative sample feature vectors can be generated to form a difficult example drug negative sample feature matrix

[0160] Finally, XR' n and XD' n are updated to obtain the difficult example drug feature matrix XR' and the difficult example disease feature matrix XD', wherein XR' and XR have the same size, and XD' and XD have the same size;

[0161] Then, the difficult example drug feature matrix FR' H and the difficult example disease feature matrix FD' H are obtained by mapping back to the hyperbolic space, as follows:

[0162] FR' H = exp o (XR')

[0163] FD' H = exp o (XD')​

[0164] where exp o Its function is to map variables to hyperbolic space. middle.

[0165] As a preferred embodiment of the drug-disease association prediction method based on hyperbolic graph feature learning network described in this invention, the specific steps of S6 are as follows:

[0166] S61: Calculate FR' H and FD' H Hyperbolic distance Distance ij ;

[0167] Calculate the feature vector of hyperbolic difficult drug cases Hyperbolic Difficulty Disease Feature Vector Hyperbolic distance Distance ij The calculation formula is as follows:

[0168]

[0169] in, Distance ij The smaller the value, the higher the likelihood of a drug-disease association;

[0170] S62: Using Distance, determine the predicted probability output of the drug-disease association through MLP;

[0171] output = MLP(Distance)

[0172] in,

[0173] Compared with existing technologies:

[0174] This invention introduces hyperbolic homogeneous fusion attention through a hyperbolic graph feature reconstructor. Based on hyperbolic space, it learns compact and representative similarity features of drug nodes and disease nodes in homogeneous graphs, thereby more accurately capturing the similarity between drug nodes and between disease nodes. This avoids the oversimplification problem that occurs when drug and disease association prediction models learn similarity features between drug nodes and between disease nodes in Euclidean space. The hyperbolic heterogeneous dynamic graph converter introduces hyperbolic heterogeneous dynamic attention, thereby more accurately capturing the semantic associations and topological relationships between drug nodes and disease nodes in heterogeneous graphs. Semantic associations and topological relationships are important components of heterogeneous features, thus generating compact and representative heterogeneous features. This avoids the distortion problem caused by drug and disease association prediction models learning heterogeneous features of drug nodes and disease nodes in heterogeneous graphs, and also avoids attention noise problems. The hyperbolic collaborative representation learning strategy fuses the similar features learned by the hyperbolic graph feature reconstructor and the heterogeneous features learned by the hyperbolic heterogeneous dynamic graph converter through MLP. This strategy enables the model to generate comprehensive and high-quality features, allowing for a more comprehensive and in-depth exploration of the association between drugs and diseases. It significantly improves the model's generalization ability and avoids problems such as the model's inability to generate comprehensive and high-quality features. The positive-negative fusion hard sample sampling strategy synthesizes the most informative negative samples, significantly improving the representativeness and information content of negative samples. This allows the model to better distinguish the boundaries between positive and negative samples, avoiding problems such as ignoring important negative samples or introducing irrelevant noisy negative samples. Attached Figure Description

[0175] Figure 1 This is a diagram showing the main structure of the hyperbolic graph feature learning network of this invention;

[0176] Figure 2 This is a schematic diagram of the process of the present invention;

[0177] Figure 3 This is a schematic diagram of the hyperbolic graph feature reconstructor structure of the present invention;

[0178] Figure 4 This is a schematic diagram of the hyperbolic heterogeneous transformation graph converter structure of the present invention;

[0179] Figure 5 This is a schematic diagram of the positive and negative fusion sampling strategy of the present invention. Detailed Implementation

[0180] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0181] This invention provides a method for predicting drug-disease associations based on hyperbolic graph feature learning networks. Please refer to [link to relevant documentation]. Figures 1-5, including specific steps as follows:

[0182] S1: Obtain drug-disease association data from the database, which is used to construct drug similarity matrix DR and disease similarity matrix DS, and map the feature extraction to hyperbolic space to obtain hyperbolic drug initial feature matrix ZR H , hyperbolic disease initial feature matrix ZD H , which is used to construct drug-disease adjacency matrix A;

[0183] The specific steps of S1 are as follows:

[0184] S11: Construct drug similarity matrix DR and disease similarity matrix DS, and obtain drug initial feature matrix ZR and disease initial feature matrix ZD through MLP;

[0185] The specific steps of S11 are as follows:

[0186] S111: Obtain drug-disease association data from the database, construct drug similarity matrix DR and disease similarity matrix DS, and obtain drug initial feature matrix ZR and disease initial feature matrix ZD through MLP; Where nr represents the number of drug nodes, and nd represents the number of disease nodes;

[0187] S112: Use MLP to extract features from DR and DS to obtain drug initial feature matrix ZR and disease initial feature matrix ZD, as follows:

[0188] ZR = MLP(DR)

[0189] ZD = MLP(DS)

[0190] Where, ZR represents the drug initial feature matrix, and d0 is the feature dimension thereof, ZR represents the i-th row of drug initial feature vector of ZR, i.e., the drug initial feature vector of drug node dr i ; ZD represents the disease initial feature matrix; ZD represents the i-th row of disease initial feature vector of ZD, i.e., the disease initial feature vector of disease node ds i ;

[0191] Where, multi-layer perceptron (MLP) is a basic neural network architecture, which is widely used in data analysis and pattern recognition tasks; it can capture the nonlinear relationship between input and output through multiple levels of node connection, so as to realize complex data modeling and learning function;

[0192] ​A homogeneous graph is a graph data structure containing only a single type of nodes and a single type of edges; the drug similarity matrix DR corresponds to a drug homogeneous graph, and the disease similarity matrix DS corresponds to a disease homogeneous graph;

[0193] Embodiment: In the experiment of the present application, the C-dataset data set is used, which consists of 663 drug nodes, 409 disease nodes, and 993 protein nodes, and has 2532 verified drug-disease association pairs, 3773 drug-protein association pairs, and disease-protein association pairs; molecular fingerprints and GIP similarity are used to calculate the similarity between drugs, if the molecular fingerprint similarity between two drugs is 0, then the GIP similarity is used as the drug similarity matrix, otherwise the average value is used as the drug similarity matrix; phenotypes and GIP similarity are used to calculate the similarity between diseases, if the phenotype similarity between two diseases is 0, then the GIP similarity is used as the disease similarity matrix, otherwise the average value is used as the disease similarity matrix; in the C-dataset data set, the size of the drug similarity matrix DR is 663x663, and the size of the disease similarity matrix DS is 409x409;

[0194] S12: Map the drug initial feature matrix ZR and the disease initial feature matrix ZD to hyperbolic space;

[0195] The specific steps of S12 are as follows:

[0196] S121: Use the Lorentz model, let K=1, then o={1,0,…,0}∈ The origin is represented as , and then is represented as

[0197] Among them, there are many equivalent hyperbolic models in hyperbolic space, such as Poincaré ball model, Lorentz model and Klein model, which exhibit different characteristics; for the Lorentz model, the constant negative curvature of the hyperbolic space is -1 / K (K>0), and the origin is the origin in the hyperbolic space , where d represents the dimension of the hyperbolic space, and K represents the constant negative curvature of the hyperbolic space, which is -1 / K (K>0);

[0198] S122: Map ZR i to hyperbolic space, the formula is as follows:

[0199]

[0200] wherein, ZR i represents the hyperbolic drug initial eigenvector of the drug node dr exp o (·) is to map the variable to the hyperbolic space (0, ZR i ) satisfies represents the Minkowski inner product; ZR i represents the drug initial eigenvector of the drug node dr i ; ||.||2 represents the L2 norm;

[0201] wherein, this is the formula of the Lorentz model mapping the eigenvector to the hyperbolic space when the constant negative curvature is-1; the Minkowski inner product is a generalized inner product definition commonly used to measure the similarity or distance in the vector space, which is a generalization of the Euclidean inner product (dot product), and in the Lorentz model, it is defined as:

[0202] S123: by performing the above operation on the drug initial eigenvector ZR i of any drug node dr i , a hyperbolic drug initial eigenvector matrix ZR

[0203] S124: repeat the above steps to obtain a hyperbolic disease initial eigenvector matrix ZD

[0204] S13: constructing a drug-disease adjacency matrix

[0205] The specific steps of S13 are as follows:

[0206] Obtaining drug-disease association data, constructing an adjacency matrix A, and the specific construction method is as follows:

[0207]

[0208] Embodiment: In the experiment of the present application, the C-dataset data set is obtained, which is composed of 663 drug nodes, 409 disease nodes, 993 protein nodes, and also has 2532 verified drug-disease association pairs, 3773 drug-protein association pairs, and disease-protein association pairs; according to the 2532 drug-disease association pairs, an adjacency matrix A is constructed;

[0209] S2: ZR H and ZD HThe hyperbolic graph feature reconstructor is applied to obtain the hyperbolic drug similarity feature matrix SR. H Similarity feature matrix of hyperbolic disease SD H ;

[0210] This invention designs a Hyperbolic Graph Feature Reconstructor (HGFR), which is a hyperbolic graph feature reconstructor based on GAT. It innovatively incorporates Hyperbolic Homogeneous Fusion Attention to extract similar features (hyperbolic similarity features) between drug nodes and disease nodes in hyperbolic space using homogeneous graphs. The steps include: extracting intermediate feature vectors between drug nodes and disease nodes using hyperbolic feature transformation; calculating the hyperbolic homogeneous fusion attention; and using the hyperbolic homogeneous fusion attention to extract the hyperbolic similarity features between drug and disease nodes. The structure of the hyperbolic graph feature reconstructor is as follows: Figure 3 As shown:

[0211] The specific steps of S2 are as follows:

[0212] S21: Use hyperbolic feature transformation to extract the intermediate feature vectors of drug nodes and disease nodes;

[0213] The specific steps of S21 are as follows:

[0214] S211: For nr drug nodes in drug-disease association data, the initial hyperbolic drug feature matrix ZR is known. H ; ZR H The initial feature vector of the hyperbolic drug in the i-th row;

[0215] At layer 0, the hyperbolic drug similarity feature matrix is ​​as follows:

[0216] SR 0,H =ZR H

[0217] but Indicates SR 0,H The hyperbolic drug similarity feature vector in the i-th row;

[0218] S212: For any drug node dr in drug-disease association data i In the hyperbolic graph feature reconstructor of layer l, a hyperbolic feature transformation is first performed to map the hyperbolic space of the previous layer to the hyperbolic space of this layer. The hyperbolic feature transformation formula for layer l is as follows:

[0219]

[0220] The feature transformation formula for hyperbolic space is specifically used to map feature vectors to the next layer of hyperbolic space; it is named the hyperbolic feature transformation here. The logarithmic mapping log... o The function of (·) is to map variables to the tangent space, i.e., the Euclidean space; the exponential mapping exp o The function of (·) is to map variables to hyperbolic space;

[0221] in, Represented as drug node dr i After hyperbolic feature transformation, located in hyperbolic space The intermediate feature vector; d are the trainable parameters of the l-th layer. l It is the dimension of the l-th layer; It is a drug node DR i Hyperbolic drug similarity feature vector at layer l-1; b l The lth layer is located at The Euclidean vectors are trainable parameters; Indicates drug node dr i The next layer, l, is located in hyperbolic space. intermediate variables in; exp o The function of (·) is to map the variable to the hyperbolic space of the next level l. log o The function of (·) is to map the variable to the tangent space of the origin o. Function is to use points Centered on the hyperbolic space, the variables are mapped to the hyperbolic space. middle; Will The tangent space from the origin o Parallel transport to destination tangent space

[0222] S22: Calculate hyperbolic homogeneous fusion attention and use it to extract hyperbolic similarity features of drug and disease nodes;

[0223] The specific steps of S22 are as follows:

[0224] S221, Calculate hyperbolic homogeneous fusion attention, which consists of node feature attention, hyperbolic distance attention, and structural attention;

[0225] S2211: Calculate the drug node dr in the l-th layer i drug node dr j Node feature attention The specific formula is as follows:

[0226]

[0227] Where LeakReLU is the activation function; These are trainable parameters; log0(·) maps the variables to the tangent space of the origin o. express and Perform matrix multiplication.

[0228] Then, normalization is performed to obtain the final node feature attention. The formula is as follows:

[0229]

[0230] in, Indicates drug node dr i The first-order neighbor set, which contains itself; when the drug node dr i and drug node dr j The higher the similarity of node features in layer l, the better. The larger;

[0231] S2212: Calculate the drug node dr in layer l. i drug node dr j hyperbolic distance attention The specific formula is as follows:

[0232]

[0233] Where δ and η are hyperparameters used to adjust the hyperbolic distance; Used for calculation and The hyperbolic distance between them; Represents the square of the hyperbolic distance;

[0234] Then, normalization is performed to obtain the final hyperbolic distance attention. The specific formula is as follows:

[0235]

[0236] in, Indicates drug node dr i The first-order neighbor set, which contains itself; when the drug node dr i and drug node dr j The smaller the hyperbolic distance at layer l, the better. The larger;

[0237] S2213: Calculate drug node dr i drug node dr jThe structural attention is calculated as follows:

[0238] Given the drug similarity matrix DR, drug node dr i The degree of dr

[0239]

[0240] The formula means summing each column j of the ith row of DR, i.e., the drug node dr i The similarity of each drug node dr j is summed up;

[0241] The drug node dr i The structural attention of the drug node dr j The calculation formula is as follows:

[0242]

[0243] Wherein,

[0244] Then the final structural attention is normalized The formula is as follows:

[0245]

[0246] Wherein, The first-order neighbor set of the drug node dr i , including itself; when the degree of the drug node dr i and the drug node dr j is smaller, the is larger;

[0247] S2214: In the lth layer, the node feature attention The hyperbolic distance attention and the structural attention are fused to obtain the hyperbolic homogeneous fusion attention, and the formula is as follows:

[0248]

[0249] Wherein w NF , w DS and w ST are trainable parameters, all initialized to 1;

[0250] S222: The hyperbolic similarity features of the drug and disease nodes are extracted using the hyperbolic homogeneous fusion attention;

[0251] S2221: The is hyperbolically encoded to obtain the drug node dr​i Hyperbolic Similarity Features in the 1st Layer The formula of hyperbolic encoding is as follows:

[0252]

[0253] wherein, is to map the variable into the hyperbolic space ; σ is an activation function; represents the first-order neighbor set of the drug node dr i , including itself; is to map the variable into the tangent space of ;

[0254] S2222: Perform the above operation on any drug node dr i , and after h-layer iteration, the hyperbolic drug similarity feature matrix ZR is obtained. represents SR h,H ;

[0255] S2223: Repeat the above steps to obtain the hyperbolic disease similarity feature matrix ZD

[0256] Embodiment: Take the hyperbolic drug initial feature matrix ZR H and the hyperbolic disease initial feature matrix ZD H obtained by S1 as input to obtain the hyperbolic drug similarity feature matrix and the hyperbolic disease similarity feature matrix in the 0th layer, ZR H contains the hyperbolic drug initial feature vector of 663 drug nodes, and the feature dimension is d0, d0 is a hyperparameter; ZD H contains the hyperbolic disease initial feature vector of 409 disease nodes, and the feature dimension is d0, d0 is a hyperparameter; in the 0th layer, the processing process of the drug node is as follows:

[0257] First step: for a drug node, use the hyperbolic feature transformation formula to map the hyperbolic drug similarity feature vector of the drug node to the next layer hyperbolic space (i.e. the 1st layer) to obtain the intermediate feature vector of the drug node, and the feature dimension is d1;

[0258] Second step: based on the intermediate feature vector of the drug node, calculate the node feature attention in the 1st layer, the hyperbolic distance attention in the 1st layer, and the structure attention (the structure attention is independent of the level), give different weights (the weight is a trainable parameter) to the three kinds of attention, and finally fuse them by weighting to form a hyperbolic homogeneous fusion attention;

[0259] ​Third step: based on the intermediate feature vector of the drug node, the hyperbolic drug similar feature vector of the drug node is updated using the hyperbolic homogeneous fusion attention, and the hyperbolic drug similar feature vector of the drug node of the first layer is obtained;

[0260] Fourth step: repeat the operation of the first three steps for each row of the hyperbolic drug similar feature matrix of the 0th layer (i.e. the hyperbolic drug similar feature vector of each drug node), and obtain the hyperbolic drug similar feature of the first layer; the remaining layers are similar to the above process, and the above process is iterated h times, and the hyperbolic drug similar feature matrix SR H , similarly, the hyperbolic disease similar feature matrix SD H can also be obtained;

[0261] S3: construct a heterogeneous graph AG according to A, and train the hyperbolic drug heterogeneous feature matrix HR H and the hyperbolic disease heterogeneous feature matrix HD H based on AG using GCN and hyperbolic heterogeneous dynamic graph transformer; H H ;

[0262] The application designs a hyperbolic heterogeneous dynamic graph transformer (HHDGT), which is a hyperbolic heterogeneous dynamic graph transformer designed based on Transformer, and innovatively introduces hyperbolic heterogeneous dynamic attention. The hyperbolic drug heterogeneous feature matrix and the hyperbolic disease heterogeneous feature matrix are trained by means of the hyperbolic drug heterogeneous initial feature matrix and the hyperbolic disease heterogeneous initial feature matrix obtained by GCN training. The process is to calculate the hyperbolic heterogeneous dynamic attention, and learn the heterogeneous features of the drug nodes and the disease nodes using the hyperbolic heterogeneous dynamic attention. The structure of the hyperbolic heterogeneous dynamic graph transformer is as shown in Figure 4 ;

[0263] The specific steps of S3 are as follows:

[0264] S31: obtain the hyperbolic disease heterogeneous initial feature matrix and the hyperbolic drug heterogeneous initial feature matrix using GCN;

[0265] The specific steps of S31 are as follows:

[0266] S311: construct a heterogeneous graph AG based on the adjacency matrix A, as follows:

[0267]

[0268] Wherein,

[0269] ​S312: In order to embody the number of layers, let MR 0,H , H MD 0,H be expressed as ZR H , ZD 0,H ; in order to use the GCN formula, splice MR 0,H , MD 0 to obtain:

[0270]

[0271] wherein, E 0,H is the feature matrix of the 0th layer spliced by MR 0,H , MD l ;

[0272] S313: In the lth layer, E l is obtained based on the heterogeneous graph AG using GCN, and the formula is as follows:

[0273]

[0274] AG'=(AG+I)

[0275]

[0276]

[0277] wherein GCN (Graph Convolutional Network) is a deep learning model for processing graph structure data; it performs information propagation and feature extraction on graph structure through convolution operation, and is widely used in node classification, graph classification, link prediction and other tasks; the heterogeneous graph is a graph structure data containing multiple types of nodes (entities) and edges (relationships); unlike the homogeneous graph (all nodes and edge types are the same), the nodes and edges in the heterogeneous graph have clear type distinction, which can more naturally model the multi-element interaction relationship in complex systems, and the heterogeneous graph AG herein has the same type nodes not connected;

[0278] wherein, is the i th row hyperbolic heterogeneous initial feature vector of E o in the lth layer; σ is an activation function; exp i (·) is used to map variables to hyperbolic space is the first-order neighbor set of the drug node dr i or the disease node ds o , including itself; is the processed adjacency matrix, is the value of the i th row and j th column; I is the unit matrix; are trainable parameters; log o (·) acts to map the variable to the tangent space of the origin o D matrix is used to calculate

[0279] S314: After h layer iterations, we get:

[0280]

[0281] That is, from E h Hyperbolic drug heterogeneity initial feature matrix and hyperbolic disease heterogeneity initial feature matrix

[0282] S32: Calculate hyperbolic heterogeneity variable attention, and use hyperbolic heterogeneity variable attention to learn the heterogeneity features of drug nodes and disease nodes;

[0283] The specific steps of S32 are as follows:

[0284] Given the hyperbolic drug heterogeneity initial feature matrix and the hyperbolic disease heterogeneity initial feature matrix, that is, MR h,H and MD h,H ;

[0285] In the 0th layer, the hyperbolic heterogeneity drug feature matrix and the hyperbolic heterogeneity disease feature matrix are respectively:

[0286] HR 0,H = MR h,H

[0287] HD 0,H = MD h,H

[0288] The hyperbolic heterogeneity variable graph transformer has three inputs: query matrix Q 1 and Q 2 , key matrix K 1 and K 2 , and value matrix V; if hyperbolic drug heterogeneity features are extracted, the query matrix is obtained through HR 0,H , and the key matrix and the value matrix are obtained through HD 0,H ; if hyperbolic disease heterogeneity features are extracted, the situation is just the opposite; in the lth layer, the outputs are HR l,H and HD l,H ;

[0289] In the lth layer, the process of obtaining HR l,H is as follows:

[0290]

[0291] where, HRl-1,H is the hyperbolic drug heterogeneous feature matrix of the l-1th layer; exp o (·) acts to map the variable to the hyperbolic space log o (·) acts to map the variable to the tangent space of the origin o represents and perform matrix multiplication, Similarly;

[0292]

[0293] where HD l-1,H is the hyperbolic disease heterogeneous feature matrix of the l-1th layer; The meaning is the same as above;

[0294] The update equation of the drug node dr i of the lth layer is as follows:

[0295]

[0296] where, represents the hyperbolic drug heterogeneous knowledge feature vector of the drug node dr i of the lth layer; V j is the jth row of V; represents the Lorentz norm; is the hyperbolic drug heterogeneous feature vector of the i-th row of HR l,H ; Norm is LayerNorm or BatchNorm; exp o (·) acts to map the variable to the hyperbolic space log o (·) acts to map the variable to the tangent space of the origin o

[0297] where BatchNorm is a deep learning technique proposed by Sergey Ioffe and Christian Szegedy in 2015 to solve the problem of gradient disappearance or gradient explosion in deep neural networks and accelerate the training process of the model; LayerNorm was proposed by Jimmy Lei Ba and Jamie Ryan Kiros et al. in 2016; Its main purpose is to help the neural network converge faster and more stably;

[0298] RDIF ij is defined as:

[0299]

[0300] where RDIF ij is the hyperbolic variable attention score, λ l is defined as:

[0301]

[0302] where λ l is the scalar of the l-th layer, is the learnable parameter of the l-th layer, λ init is the hyperparameter;

[0303] After h layers of iteration, the drug node dr i The output hyperbolic drug heterogeneous feature vector is and the hyperbolic drug heterogeneous feature matrix is HR H represents HR h,H ;

[0304] Similarly, the hyperbolic disease heterogeneous feature matrix HD H can also be obtained at the h-th layer.

[0305] S4: splice SR H and HR H , splice SD H and HD H , and use MLP to obtain the hyperbolic drug feature matrix FR H and the hyperbolic disease feature matrix FD H ;

[0306] The specific steps of the S4 are as follows:

[0307] The hyperbolic drug and disease feature matrices are calculated, and the formula is as follows:

[0308] FR H = exp o (MLP(log o (SR H )||log o (HR H )))

[0309] FD H = exp o (MLP(log o (SD H )||log o (HD H )))

[0310] where exp o (·) is used to map the variable to the hyperbolic space log o The function is to map the variable to the tangent space of the origin o ; the function of MLP is to reduce the feature vector of hyperbolic drug and disease from to

[0311] S5: based on FR H and FD H , a positive and negative fusion difficult example sampling strategy is used to construct hyperbolic difficult example drug feature matrix FR' H and hyperbolic difficult example disease feature matrix FD' H ;

[0312] The application designs a positive and negative fusion difficult example sampling strategy, which is used to generate difficult example drug negative samples and difficult example disease negative samples to improve the decision boundary of the model, and the process includes constructing drug negative samples and disease negative samples; using the positive and negative fusion difficult example sampling strategy to generate difficult example drug negative samples and difficult example disease negative samples and obtaining hyperbolic difficult example drug feature matrix and hyperbolic difficult example disease feature matrix; the structure of the positive and negative fusion difficult example sampling strategy is as shown in Figure 5 :

[0313] The specific steps of S5 are as follows:

[0314] S51: constructing drug negative samples and disease negative samples;

[0315] The specific steps of S51 are as follows:

[0316] S511: first map the hyperbolic drug feature matrix and the hyperbolic disease feature matrix to the drug feature matrix XR and the disease feature matrix XD as follows:

[0317] XR=log o (FR H )

[0318] XD=log o (FD H )

[0319] Wherein, log o The function is to map the variable to the tangent space of the origin o ;

[0320] S512: Obtain drug-disease association data, generate a heterogeneous graph G, G contains drug-disease relationships, drug-drug relationships and disease-disease relationships; find all positive sample pairs and negative sample pairs from the heterogeneous graph G, first ensure that the number of positive sample pairs and negative sample pairs is the same, the strategy adopted is: randomly select a part of negative sample pairs from the heterogeneous graph G to ensure that the number is the same as the positive sample pairs, and vice versa; finally, select drug negative samples and disease negative samples from the screened negative sample pairs; wherein the drug negative sample feature matrix is The disease negative sample feature matrix is N neg is the number of negative sample pairs, that is, the number of drug negative samples and disease negative samples;

[0321] Wherein, the heterogeneous graph is a graph structure data containing multiple types of nodes (entities) and edges (relationships); Unlike homogeneous graphs (all nodes and edge types are the same), the nodes and edges in the heterogeneous graph have clear type distinction, which can more naturally model the multi-element interaction relationship in the complex system, and the same type nodes in the heterogeneous graph G may be connected;

[0322] Embodiment: In the experiment of the present application, C-dataset data set is obtained, which consists of 663 drug nodes and 409 disease nodes, and the adjacency matrix A is obtained from S1, and 2532 drug-disease association pairs are obtained from the adjacency matrix, thereby obtaining drug-disease relationships;

[0323] The drug similarity matrix and the disease similarity matrix are obtained in S1, the drug-drug relationship is obtained according to the drug similarity matrix, and the disease-disease relationship is obtained according to the disease similarity matrix;

[0324] The heterogeneous graph G is constructed according to the drug-drug relationship, the disease-disease relationship and the drug-disease relationship, the heterogeneous graph G has 663 drug nodes and 409 disease nodes; There are 2532 verified drug-disease association pairs, so the number of positive sample pairs is 2532, and the number of negative sample pairs is 268635 (663*409-2532);

[0325] In the 268635 negative sample pairs, 2532 negative sample pairs are randomly selected, so that the number of positive sample pairs and negative sample pairs is consistent, both of which are 2532;

[0326] S52: Generate difficult example drug negative samples and difficult example disease negative samples using positive-negative fusion difficult example sampling strategy, and obtain hyperbolic difficult example drug feature matrix and hyperbolic difficult example disease feature matrix;

[0327] The specific steps of S52 are as follows:

[0328] S521: Construct difficult example candidate negative sample feature matrix;

[0329] For any node in a heterogeneous graph G, the set of all first-order and second-order neighbor nodes is called the neighbor set of that node.

[0330] For all negative drug samples, find all their first-order and second-order neighbors, i.e., the neighbor set;

[0331] For any negative drug sample r, the operation to construct the feature matrix of hard example candidate negative samples is as follows:

[0332] First, based on the heterogeneous graph G, find all positive drug samples in its neighbor set. The local neighbor positive drug sample feature matrix formed by these positive drug samples is: XR pl Then, the feature vectors of all positive drug samples are fused to obtain the fused positive drug sample feature vector x. pr The formula is as follows:

[0333]

[0334] in It's XR pl The mr-th row is the feature vector of the positive drug sample; mr represents the number of positive drug samples in the neighbor set of the negative drug sample r.

[0335] Similarly, the feature vectors of all positive disease samples in its neighbor set are fused to obtain the fused positive disease sample feature vector x. pd The formula is as follows:

[0336]

[0337] in Yes XD pl The md-th row is the feature vector of the positive disease sample; md represents the number of positive disease samples in the neighbor set of the negative drug sample r.

[0338] Then, find the disease negative samples of the second-order neighbors within the neighbor set of the drug negative sample r. The resulting second-order neighbor disease negative sample feature matrix is: XD nl According to x pr and x pd The probability distribution of disease negative samples is calculated for the second-order neighbors of the disease negative samples within the neighbor set of the drug negative sample r, using the following formula:

[0339]

[0340] in, XD nlthe qth row of the feature vector of the drug; μ is a hyperparameter, used to balance the influence of drugs and diseases; N represents the number of disease-negative samples in the second-order neighbors of the drug-negative sample r; the meaning of this formula is that the closer the distance (x pr ,x pd ), the greater the probability of being selected as a disease-negative sample;

[0341] Then, according to the disease-negative sample probability distribution , the disease-negative samples in the second-order neighbors of the drug-negative sample r are selected, and a candidate disease-negative sample feature matrix T of size M is constructed M is a hyperparameter;

[0342] Next, the information of the disease fusion positive sample feature vector x pd is injected into the candidate disease-negative sample feature matrix T to construct a hard example candidate disease-negative sample feature matrix T', and the formula is as follows:

[0343] T' i = αx pd + (1-α)T i , α∈(0,1)

[0344] Where T' i is the ith row of the hard example disease-negative sample feature vector of T'; α is a hyperparameter; T i is the ith row of the candidate disease-negative sample feature vector of T;

[0345] S522: Obtain a hyperbolic hard example drug feature matrix and a hyperbolic hard example disease feature matrix;

[0346] For each row of the hard example disease-negative sample feature vector T' j of the hard example candidate disease-negative sample feature matrix T', the dot product is taken with the drug-negative sample feature vector of the drug-negative sample r, and the formula is as follows:

[0347]

[0348] Where, is the hard example disease-negative sample feature vector with the largest dot product selected from T';

[0349] Since there are N neg drug-negative samples, N neg hard example disease-negative sample feature vectors can be generated to form a hard example disease-negative sample feature matrix

[0350] Similarly, for disease-negative samples, N neg hard example drug-negative sample feature vectors can be generated to form a hard example drug-negative sample feature matrix

[0351] Finally, based on XR' n and XD' n update XR and XD, and then get the difficult example drug feature matrix XR' and the difficult example disease feature matrix XD', wherein XR' and XR have the same size, and XD' and XD have the same size;

[0352] Then map back to the hyperbolic space to get the hyperbolic difficult example drug feature matrix FR' H and the hyperbolic difficult example disease feature matrix FD' H , as follows:

[0353] FR' H = exp o (XR')

[0354] FD' H = exp o (XD')

[0355] Where exp o acts to map the variable to the hyperbolic space ;

[0356] Embodiment: 2532 negative sample pairs screened out from the S51 embodiment are respectively screened out 2532 drug negative samples and 2532 disease negative samples, and then the positive-negative fusion difficult example sampling strategy is applied to the drug negative samples and the disease negative samples respectively; the (drug, disease) node pair not connected by any known association is called a negative sample pair, for example, there are 3 drug nodes and 3 disease nodes, drug node 1 and disease node 1 are not connected, then drug node 1 and disease node 1 are negative sample pairs, and drug node 1 and disease node 1 are respectively called drug negative sample 1 and disease negative sample 1; the negative sample is local, for example, drug node 1 is a drug negative sample, but it can still be connected with other disease nodes;

[0357] The process of applying the positive-negative fusion difficult example sampling strategy to the drug negative sample is as follows: for any drug negative sample r, the first step is to find all its first-order and second-order neighbors, i.e. the neighbor set, then fuse the feature vectors of all drug positive samples in the neighbor set (as long as there is a disease node connected with it, it is a drug positive sample) to get the drug fusion positive sample feature vector x pr fuse the feature vectors of all disease positive samples in the neighbor set to get the disease fusion positive sample feature vector x pd (as long as there is a drug node connected with it, it is a disease positive sample);

[0358] Second step: find all disease negative samples of second-order neighbors (all disease nodes of second-order neighbors that are not connected to drug negative sample node r), and then calculate disease negative sample probability distribution according to x pr and x pd , and select M disease negative samples with the maximum disease negative sample probability values according to the disease negative sample probability distribution;

[0359] Third step: integrate x pd into M disease negative samples to obtain M hard example disease negative samples, and the feature vector of the hard example disease negative sample after integrating x pd is called a hard example disease negative sample feature vector;

[0360] Fourth step: perform dot product operation on the hard example disease negative sample feature vector of each of the M hard example disease negative samples and the feature vector of drug negative sample r, and select the hard example disease negative sample with the maximum dot product value as the final result, while eliminating the remaining hard example disease negative samples (the feature vector of the eliminated hard example disease negative sample is not updated to the hard example disease negative sample feature vector);

[0361] Fifth step: 2532 drug negative samples obtain 2532 hard example disease negative samples, and obtain a hard example disease negative sample feature matrix XD' n , and then update the disease feature matrix XD to obtain a hard example disease feature matrix XD', and map it back to the hyperbolic space to obtain a hyperbolic hard example disease feature matrix FD' H ; if multiple drug negative samples select the same hard example disease negative sample (the hard example disease negative sample has multiple hard example disease negative sample feature vectors), a hard example disease negative sample feature vector is randomly selected as the final hard example disease negative sample feature vector of the hard example disease negative sample;

[0362] Similarly, the positive-negative fusion hard example sampling strategy can also be used for disease negative samples to obtain a hard example drug negative sample feature matrix XR' n , and then update the drug feature matrix XR to obtain a hard example drug feature matrix XR', and map it back to the hyperbolic space to obtain a hyperbolic hard example drug feature matrix FR' H ;

[0363] S6: use FR' H and FD' H to output a prediction result through an MLP;

[0364] The specific steps of S6 are as follows:

[0365] S61: calculate the hyperbolic distance Distance ij between FR' H and FD' H ;

[0366] Hyperbolic difficult case drug feature vector is calculated and hyperbolic difficult case disease feature vector The hyperbolic distance Distance ij , the calculation formula is as follows:

[0367]

[0368] Wherein, Distance ij The smaller, the higher the possibility of drug-disease association;

[0369] S62: using Distance, judging the prediction association probability output of drug-disease by MLP;

[0370] output=MLP(Distance)

[0371] Wherein,

[0372] Based on the above, the technical purposes of the present application are:

[0373] 1. In view of the problem that the drug and disease association prediction model simplifies too much when learning similar features in homogeneous graph in Euclidean space, a hyperbolic graph feature reconstructor is proposed. The hyperbolic graph feature reconstructor introduces hyperbolic homogeneous fusion attention to learn compact and discriminative similar features of drug nodes and disease nodes in hyperbolic space through homogeneous graph, so as to more accurately capture the similarity between drug nodes and drug nodes, and between disease nodes and disease nodes.

[0374] 2. In view of the problem that the drug and disease association prediction model is prone to distortion when learning heterogeneous features in heterogeneous graph in Euclidean space, a hyperbolic heterogeneous variable graph converter is proposed. Through the designed hyperbolic heterogeneous variable attention, the hyperbolic heterogeneous variable graph converter can learn the semantic association and topological relationship of the heterogeneous graph in the hyperbolic space, and the semantic association and topological relationship are important components of the heterogeneous features, so it can generate compact and representative heterogeneous features, thereby accurately modeling the complex relationship between drug nodes and disease nodes, not only solving the distortion problem, but also avoiding the attention noise problem.

[0375] 3. To address the problem of limited feature representation capability, the model cannot generate comprehensive and high-quality features, a hyperbolic collaborative representation learning strategy is proposed. The hyperbolic graph feature reconstructor learns the similar features of drug nodes and disease nodes in the homogeneous graph, the hyperbolic heterogeneous variation graph converter learns the heterogeneous features of drug nodes and disease nodes in the heterogeneous graph, and finally the similar features and heterogeneous features are fused. Through this strategy, the model can generate comprehensive and high-quality features, so as to more comprehensively and deeply explore the association between drug nodes and disease nodes, and significantly improve the generalization ability of the model.

[0376] 4. To address the problem of how to select the most informative negative samples, a positive-negative fusion difficult example sampling strategy is proposed to synthesize difficult drug negative samples and difficult disease negative samples. The positive-negative fusion difficult example sampling strategy synthesizes the most informative negative samples through negative sample probability distribution and fusion positive samples. This strategy can help the model better distinguish positive and negative sample pairs, thereby improving the prediction performance of the model.

[0377] Although the present application has been described above with reference to the implementation, various improvements can be made thereto and equivalents can be substituted without departing from the scope of the present application. In particular, features of the disclosed implementations can be combined together in any manner, provided that there is no structural conflict. Since such combinations are not exhaustively described in the specification, they are not described herein only for the sake of brevity and resource conservation. Therefore, the present application is not limited to the specific implementations disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting drug-disease associations based on hyperbolic graph feature learning networks, characterized in that, The specific steps are as follows: S1: Obtain drug-disease association data from the database to construct the drug similarity matrix DR and the disease similarity matrix DS. After feature extraction, map the data into hyperbolic space to obtain the hyperbolic initial drug feature matrix ZR. H Hyperbolic disease initial feature matrix ZD H , used to construct the drug-disease adjacency matrix A; S2: Against ZR H and ZD H The hyperbolic graph feature reconstructor is applied to obtain the hyperbolic drug similarity feature matrix SR. H Similarity feature matrix of hyperbolic disease SD H The specific steps are as follows: S21: Extract the intermediate feature vectors of drug nodes and disease nodes using hyperbolic feature transformation; the specific steps are as follows: S211: For nr drug nodes in drug-disease association data, the initial hyperbolic drug feature matrix ZR is known. H , ZR H The initial feature vector of the hyperbolic drug in the i-th row; S212: For any drug node dr in drug-disease association data i In the hyperbolic graph feature reconstructor of the l-th layer, a hyperbolic feature transformation is first performed to map the hyperbolic space of the previous layer to the hyperbolic space of this layer. S22: Calculate hyperbolic homogeneous fusion attention, and use hyperbolic homogeneous fusion attention to extract hyperbolic similarity features of drug and disease nodes. The specific steps are as follows: S221, Calculate hyperbolic homogeneous fusion attention, which consists of node feature attention, hyperbolic distance attention and structural attention; S2211: Calculate the drug node dr in the l-th layer i drug node dr j Node feature attention S2212: Calculate the drug node dr in layer l. i drug node dr j hyperbolic distance attention S2213: Calculate drug node dr i drug node dr j Structural attention; S2214: In the l-th layer, pay attention to the node features. Hyperbolic distance attention and structural attention Fusion yields hyperbolic homogeneous fusion attention; S222: Using hyperbolic homogeneous fusion attention to extract hyperbolic similarity features of drug and disease nodes; S2221: Yes Perform hyperbolic encoding to obtain the drug node dr i Hyperbolic similarity features at layer l S2222: For any drug node dr i After performing the above operations and iterating through h layers, the hyperbolic drug similarity feature matrix can finally be obtained. use Indicates SR h,H ; S2223: Repeat the above steps to obtain the hyperbolic disease similarity feature matrix. S3: Construct a heterogeneous graph AG based on A, and then use GCN and a hyperbolic heterogeneous graph converter to process the ZR graph based on AG. H and ZD H Training yields the hyperbolic drug heterogeneity feature matrix HR H Hyperbolic Disease Heterogeneous Feature Matrix (HD) H ; S4: SR splicing H and HR H splicing SD H and HD H The hyperbolic drug feature matrix FR was obtained using MLP. H and hyperbolic disease feature matrix FD H ; S5: Based on FR H and FD H A hyperbolic hard case drug feature matrix FR' was constructed using a positive-negative fusion hard case sampling strategy. H Hyperbolic Difficult Case Disease Feature Matrix FD' H ; S6: Utilizing FR' H and FD' H The prediction results are output through MLP.

2. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Construct the drug similarity matrix DR and the disease similarity matrix DS, and obtain the initial drug feature matrix ZR and the initial disease feature matrix ZD through MLP; S12: Map the initial feature matrix of the drug ZR and the initial feature matrix of the disease ZD to hyperbolic space; S13: Constructing a drug-disease adjacency matrix Where nr represents the number of drug nodes and nd represents the number of disease nodes.

3. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 2, characterized in that, The specific steps of S11 are as follows: S111: Obtain drug-disease association data from the database and construct a drug similarity matrix. Disease similarity matrix Where nr represents the number of drug nodes and nd represents the number of disease nodes; S112: Use MLP to extract features from DR and DS to obtain the initial drug feature matrix ZR and the initial disease feature matrix ZD, as shown in the following formula: ZR = MLP(DR) ZD = MLP(DS) in, This represents the initial feature matrix of the drug, where d0 is its feature dimension. This represents the initial feature vector of the drug in the i-th row of ZR, i.e., the drug node dr. i The initial feature vector of the drug; Represents the initial feature matrix of the disease; This represents the initial feature vector of the disease in the i-th row of ZD, i.e., the disease node ds. i The initial feature vector of the disease; The specific steps of S12 are as follows: S121: Using the Lorentz model, let K = 1, then Represented as The origin, then use express S122: For ZR i Mapping this onto hyperbolic space, the formula is as follows: in, Indicates drug node dr i The initial eigenvector of the hyperbolic drug. exp o The function of (·) is to map variables to hyperbolic space. (0, ZR) i )satisfy Represents the Minkowski inner product; ZR i Indicates drug node dr i The initial feature vector of the drug; ||.||2 represents the L2 norm; S123: By analyzing any drug node dr i The initial feature vector ZR of the drug i By performing the above operations, the initial feature matrix of the hyperbolic drug can be obtained. S124: Repeat the above steps to obtain the initial feature matrix of hyperbolic disease. The specific steps of S13 are as follows: Obtain drug-disease association data and construct an adjacency matrix A. The specific construction method is as follows:

4. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 1, characterized in that: At layer 0, the hyperbolic drug similarity feature matrix is ​​as follows: SR 0,H =ZR H but Indicates SR 0,H The hyperbolic drug similarity feature vector in the i-th row; The hyperbolic feature transformation formula for the l-th layer is as follows: in, Represented as drug node dr i After hyperbolic feature transformation, located in hyperbolic space The intermediate feature vector; d are the trainable parameters of the l-th layer. l It is the dimension of the l-th layer; It is a drug node DR i Hyperbolic drug similarity feature vector at layer l-1; b l The lth layer is located at The Euclidean vectors are trainable parameters; Indicates drug node dr i The next layer, l, is located in hyperbolic space. intermediate variables in; exp o The function of (·) is to map the variable to the hyperbolic space of the next level l. log o The function of (·) is to map the variable to the tangent space of the origin o. Function is to use points Centered on the hyperbolic space, the variables are mapped to the hyperbolic space. middle; b l The tangent space from the origin o Parallel transport to destination tangent space The specific formula is as follows: LeakReLU is the activation function. These are trainable parameters; log o The function of (·) is to map the variable to the tangent space of the origin o. express and Perform matrix multiplication; Then, normalization is performed to obtain the final node feature attention. The formula is as follows: in, Indicates drug node dr i The first-order neighbor set, which contains itself; when the drug node dr i and drug node dr j The higher the similarity of node features in layer l, the better. The larger; The specific formula is as follows: Where δ and η are hyperparameters used to adjust the hyperbolic distance; Used for calculation and The hyperbolic distance between them; Represents the square of the hyperbolic distance; Then, normalization is performed to obtain the final hyperbolic distance attention. The specific formula is as follows: in, Indicates drug node dr i The first-order neighbor set, which contains itself; when the drug node dr i and drug node dr j The smaller the hyperbolic distance at layer l, the better. The larger; The specific calculation process is as follows: Given the drug similarity matrix DR, and the drug node dr i The degree is defined here as: This formula means summing the values ​​of each column j in the i-th row of DR, i.e., the drug node dr. i For each drug node dr j Sum the similarities; Drug node dr i drug node dr j Structural attention The calculation formula is as follows: in, Then, normalization is performed to obtain the final structural attention. The formula is as follows: in, Indicates drug node dr i The first-order neighbor set, which contains itself; when the drug node dr i and drug node dr j The smaller the degree, the better. The larger; The formula is as follows: Where w NF w DS and w ST These are trainable parameters, all initialized to 1; The formula for hyperbolic encoding is as follows: in, Its function is to map variables to hyperbolic space. In the middle; σ is the activation function; Indicates drug node dr i The set of first-order neighbors, including itself; Its function is to map variables to tangent space middle.

5. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 1, characterized in that, The specific steps of S3 are as follows: S31: Use GCN to obtain the initial feature matrix of hyperbolic disease heterogeneity and the initial feature matrix of hyperbolic drug heterogeneity; S32: Calculate hyperbolic heterogeneous variable attention and use it to learn the heterogeneous features of drug nodes and disease nodes.

6. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 5, characterized in that, The specific steps of S31 are as follows: S311: Construct a heterogeneous graph AG based on the adjacency matrix A, as follows: in, S312: To reflect the number of layers, let MR 0,H Represented as ZR H MD 0,H Represented as ZD H In order to use the GCN formula, MR 0,H MD 0,H By splicing them together, we get: in, E 0 It is layer 0 made by MR 0,H MD 0,H The feature matrix formed by splicing together; S313: At layer l, E is obtained using GCN based on the heterogeneous graph AG. l The formula is as follows: AG' = (AG + I) in, It is E l The hyperbolic heterogeneous initial feature vector in the i-th row of the l-th layer; σ is the activation function; exp o The function of (·) is to map variables to hyperbolic space. It is a drug node DR i or disease node ds i The set of its first-order neighbors, including itself; It is the processed adjacency matrix. yes The value in the i-th row and j-th column; I is the identity matrix; These are trainable parameters; log o The function of (·) is to map the variable to the tangent space of the origin o. The D matrix is ​​used for calculation S314: After h-level iterations, we can obtain: That is, from E h The initial feature matrix of hyperbolic drug heterogeneity was obtained. and hyperbolic disease heterogeneous initial feature matrix The specific steps of S32 are as follows: Given the initial characteristic matrix of hyperbolic drug heterogeneity and the initial characteristic matrix of hyperbolic disease heterogeneity, i.e., MR h,H and MD h,H ; At level 0, the hyperbolic drug heterogeneity feature matrix and the hyperbolic disease heterogeneity feature matrix are as follows: HR 0,H =MR h,H HD 0,H =MD h,H The hyperbolic heterogeneous transformation graph converter has three inputs: the query matrix Q. 1 and Q 2 Key matrix K 1 and K 2 And the value matrix V; if extracting hyperbolic drug heterogeneity features, the query matrix is ​​obtained through HR. 0,H The key matrix and value matrix are obtained through HD. 0,H We obtain; if we extract the heterogeneous features of hyperbolic diseases, the situation is exactly the opposite; at the l-th layer, the outputs are HR l,H and HD l,H ; At level l, HR is obtained. l,H The process is as follows: Among them, HR l-1,H It is the hyperbolic drug heterogeneity feature matrix of the (l-1)th layer; exp o The function of (·) is to map variables to hyperbolic space. log o The function of (·) is to map the variable to the tangent space of the origin o. Represents log o (HR l-1,H )and Perform matrix multiplication. Similarly; Among them, HD l-1,H It is the hyperbolic disease heterogeneity feature matrix of the (l-1)th layer; Define the drug node dr in layer l. i The update equation is as follows: in, The drug node dr in layer l represents i Hyperbolic drug heterogeneous knowledge feature vector; V j It is the j-th row of V; Denotes the Lorentz norm; HR l,H The hyperbolic drug heterogeneous feature vector in the i-th row; Norm is either LayerNorm or BatchNorm; exp o The function of (·) is to map variables to hyperbolic space. log o The function of (·) is to map the variable to the tangent space of the origin o. RDIF ij Defined as: RDIF ij Let λ be the hyperbolic variable attention score. l Defined as: Where, λ l It is a scalar of the l-th layer. λ is the learnable parameter of the l-th layer. init It's a hyperparameter; After h-layer iteration, the drug node dr i The hyperbolic drug heterogeneity feature vector is This leads to the hyperbolic drug heterogeneity feature matrix. Use HR H Indicates HR h,H ; Similarly, the hyperbolic disease heterogeneity feature matrix HD can also be obtained at the h-th layer. H .

7. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 1, characterized in that, The specific steps of S4 are as follows: The hyperbolic drug and disease feature matrix is ​​calculated using the following formula: FR H =exp o (MLP(log o (SR H )||log o (HR H ))) FD H =exp o (MLP(log o (SD H )||log o (HD H ))) Where, exp o The function of (·) is to map variables to hyperbolic space. log o The function of (·) is to map the variable to the tangent space of the origin o. In the middle; the role of MLP is to extract the feature vectors of hyperbolic drugs and diseases from Dimensional reduction 8. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 1, characterized in that, The specific steps of S5 are as follows: S51: Construct negative samples for drugs and negative samples for diseases; S52: Use the positive-negative fusion hard sample sampling strategy to generate hard sample drug negative samples and hard sample disease negative samples, and obtain hyperbolic hard sample drug feature matrix and hyperbolic hard sample disease feature matrix; The specific steps of S51 are as follows: 511: First, analyze the hyperbolic drug feature matrix. and hyperbolic disease feature matrix The mapping is as follows: (Diagram showing drug feature matrix XR and disease feature matrix XD) XR=log o (FR H ) XD=log o (FD H ) in, log o Its function is to map variables to the tangent space of the origin o. middle; S512: Obtain drug-disease association data and generate a heterogeneous graph G, which includes drug-disease relationships, drug-drug relationships, and disease-disease relationships. Find all positive and negative sample pairs from the heterogeneous graph G. First, ensure that the number of positive and negative sample pairs is equal. The strategy is to randomly select a portion of negative sample pairs from the heterogeneous graph G to ensure the number is the same as the number of positive sample pairs, and vice versa. Finally, select drug negative samples and disease negative samples from the selected negative sample pairs. The feature matrix of the drug negative samples is as follows: The feature matrix of negative disease samples is N neg This represents the number of negative sample pairs, specifically the number of negative drug samples and negative disease samples. The specific steps of S52 are as follows: S521: Construct the feature matrix of candidate negative samples for difficult examples; For any node in a heterogeneous graph G, the set of all first-order and second-order neighbor nodes is called the neighbor set of that node. For all negative drug samples, find all their first-order and second-order neighbors, i.e., the neighbor set; For any negative drug sample r, the operation to construct the feature matrix of hard example candidate negative samples is as follows: First, based on the heterogeneous graph G, find all positive drug samples in its neighbor set. The local neighbor positive drug sample feature matrix formed by these positive drug samples is: XR pl Then, the feature vectors of all positive drug samples are fused to obtain the fused positive drug sample feature vector x. pr The formula is as follows: in It's XR pl The mr-th row is the feature vector of the positive drug sample; mr represents the number of positive drug samples in the neighbor set of the negative drug sample r. Similarly, the feature vectors of all positive disease samples in its neighbor set are fused to obtain the fused positive disease sample feature vector x. pd The formula is as follows: in Yes XD pl The md-th row is the feature vector of the positive disease sample; md represents the number of positive disease samples in the neighbor set of the negative drug sample r. Then, find the disease negative samples of the second-order neighbors within the neighbor set of the drug negative sample r. The resulting second-order neighbor disease negative sample feature matrix is: XD nl According to x pr and x pd The probability distribution of disease negative samples is calculated for the second-order neighbors of the disease negative samples within the neighbor set of the drug negative sample r, using the following formula: in, XD nl The q-th row of the feature vector; μ is a hyperparameter used to balance the effects of drugs and diseases; N represents the number of disease negative samples that are second-order neighbors of the drug negative sample r; the meaning of this formula is the distance (x pr ,x pd The closer the negative sample of the disease is, the greater the probability that it will be selected as a candidate negative sample of the disease. Then, based on the disease negative sample probability distribution Select the disease negative samples from the second-order neighbors of the M drug negative sample neighbor sets corresponding to the M largest probability values, and construct a candidate disease negative sample feature matrix of size M. M is a hyperparameter; Next, the disease will be fused with the positive sample feature vector x. pd The information is injected into the negative sample feature matrix T of the candidate disease to construct the negative sample feature matrix T' of the difficult case candidate disease, as shown in the following formula: T' i =αx pd +(1-α)T i ,α∈(0,1) Among them, T' i T is the feature vector of the i-th row of the difficult disease negative samples; α is the hyperparameter; T i It is the feature vector of the negative sample of the candidate disease in the i-th row of T; S522: Obtain the drug feature matrix and disease feature matrix for hyperbolic difficult cases; For each row of the feature matrix T' of the candidate disease negative samples, the feature vector T' of the candidate disease negative samples is... j The feature vector of the negative drug sample r The formula for the dot product is as follows: in, It is the feature vector of the difficult disease negative sample with the largest dot product selected from T'; Because there are N negative samples for drugs neg Therefore, N can be generated. neg The feature vectors of the negative samples of difficult-to-diagnose diseases constitute the feature matrix of the negative samples of difficult-to-diagnose diseases. Similarly, for negative disease samples, N can be generated. neg The feature vectors of the difficult-case drug negative samples constitute the feature matrix of the difficult-case drug negative samples. Finally, XR and XD are updated based on XR'n and XD'n to obtain the difficult case drug feature matrix XR' and the difficult case disease feature matrix XD', where XR' and XR have the same size, and XD' and XD have the same size. Then, by mapping back to hyperbolic space, we obtain the hyperbolic difficult-to-define drug feature matrix FR'. H Hyperbolic Difficult Case Disease Feature Matrix FD' H ,as follows: FR' H =exp o (XR') FD' H =exp o (XD') where exp o Its function is to map variables to hyperbolic space. middle.

9. The method for predicting drug-disease association based on hyperbolic graph feature learning network according to claim 1, characterized in that, The specific steps of S6 are as follows: S61: Calculate FR' H and FD' H Hyperbolic distance Distance ij ; Calculate the feature vector of hyperbolic difficult drug cases Hyperbolic Difficulty Disease Feature Vector Hyperbolic distance Distance ij The calculation formula is as follows: in, Distance ij The smaller the value, the higher the likelihood of a drug-disease association; S62: Using Distance, determine the predicted probability output of the drug-disease association through MLP; output = MLP(Distance) in,

Images